Wind energy is of great importance for future energy development. In order to fully exploit wind energy, wind farms are often located at high latitudes, a practice that is accompanied by a high risk of icing. Traditional blade icing detection methods are usually based on manual inspection or external sensors/tools, but these techniques are limited by human expertise and additional costs. Model-based methods are highly dependent on prior domain knowledge and prone to misinterpretation. Data-driven approaches can offer promising solutions but require a massive amount of labeled training data, which are not generally available. In addition, the data collected for icing detection tend to be imbalanced because, most of the time, wind turbines operate under normal conditions. To address these challenges, this article presents a novel deep class-imbalanced semisupervised (DCISS) model for estimating blade icing conditions. DCISS integrates class-imbalanced and semisupervised learning (SSL) using a prototypical network that can rebalance features and measure the similarities between labeled and unlabeled samples. In addition, a channel calibration attention module is proposed to improve the ability to extract features from raw data. The proposed model has been evaluated using the blade icing datasets of three wind turbines. Compared to the classical anomaly detection and state-of-the-art SSL algorithms, DCISS shows significant advantages in terms of accuracy. Compared to five different class-imbalanced loss functions, the proposed DCISS is competitive. The generalization and practicability of the proposed model are further verified in the use case of online estimation.

Wind farms are usually located in high-latitude areas, which bring a high risk of icing. Traditional methods of anti-blade-icing are limited by extra costs and potential damages to the original mechanical structure. Model-based methods are heavily dependent on mathematical models of the blade icing, which are prone to produce erroneous estimation. As data-driven models are better able to achieve competitive performances for the blade icing estimation, this article proposes a temporal attention-based convolutional neural network (TACNN). This novel data-driven model introduces a temporal attention module into a convolutional neural network, with the goal of determining the importance of sensors and timesteps and automatically identifying discriminative features from raw sensor data. Benchmark experiments on ten public datasets of multivariate time-series classification show competitive performance against the state-of-the-art methods. Compared with ten baseline networks and three widely used attention mechanisms, the TACNN shows significant advantages applying to three real-world datasets. These datasets are logged by the supervisory control and data acquisition system and contain operational and environmental measurements such as power and temperature. The ablation study and sensitivity study demonstrate the effectiveness of the key components of the TACNN. The practicability of the TACNN is further verified through online estimation testing.

Terahertz (THz) spectroscopy is now achieving increasing attention in security inspection owning to its non-destructiveness and deep penetrability of most packaging materials, such as leather, wood and wrapper. However, two major obstacles remain in spectral classification of liquid contraband: the complex components in some contraband and the spectral overlapping effect in similar types of contraband. In this paper, we establish two THz classification datasets and propose a real-time multi-class and multi-concentration liquid contraband spectral classification framework based on a convolutional neural network (CNN). The framework can not only identify contraband with complex components but also classify different concentrations of contraband. We also evaluate the robustness of the framework in different signal-to-noise ratio (SNR). Experimental results demonstrate that our algorithm (CNN) achieves the best performance compared with other deep learning and machine learning algorithms. It can be concluded that THz spectroscopy together with CNN is a promising technique for the classification of liquid contraband.

Semantic segmentation is of great importance and a challenge in computer vision. One of its main problems is how to efficiently obtain rich information (geometric structure) and identify useful features from higher dimensions. A light field camera, due to its special microlens array structure, can completely record the angular-spatial information of scenes, which is attractive and has great potential to improve the performance of semantic segmentation tasks. Inspired by this, we propose an end-to-end semantic segmentation network that can process light field macropixel images robustly and extract their features. In addition, this network can flexibly and efficiently load the different popular deep learning backbones. Furthermore, we propose an efficient angular model, which, to learn the angular features between the different viewpoints of the macropixel image, improves the nonlinearity of angular-spatial features and enhances multichannel semantic correlations. To evaluate the network, we construct a new real scene light field dataset comprising 800 high-quality samples. The quantitative and qualitative results show that the highest mean intersection over union (mIoU) based on our algorithm is greater than 57%. Our algorithm achieves a 10.30% increase compared with state-of-the-art semantic segmentation algorithms. In combination with different backbones or multiscale light field macropixel images, the network can also achieve comparable results. This preliminary work demonstrates that the combination of light field imaging and deep learning technology has potential applications in the future study of semantic segmentation.

Deep Learning-based chest Computed Tomography (CT) analysis has been proven to be effective and efficient for COVID-19 diagnosis. Existing deep learning approaches heavily rely on large labeled data sets, which are difficult to acquire in this pandemic situation. Therefore, weakly-supervised approaches are in demand. In this paper, we propose an end-to-end weakly-supervised COVID-19 detection approach, ResNext+, that only requires volume level data labels and can provide slice level prediction. The proposed approach incorporates a lung segmentation mask as well as spatial and channel attention to extract spatial features. Besides, Long Short Term Memory (LSTM) is utilized to acquire the axial dependency of the slices. Moreover, a slice attention module is applied before the final fully connected layer to generate the slice level prediction without additional supervision. An ablation study is conducted to show the efficiency of the attention blocks and the segmentation mask block. Experimental results, obtained from publicly available datasets, show a precision of 81.9% and F1 score of 81.4%. The closest state-of-the-art gives 76.7% precision and 78.8% F1 score. The 5% improvement in precision and 3% in the F1 score demonstrate the effectiveness of the proposed method. It is worth noticing that, applying image enhancement approaches do not improve the performance of the proposed method, sometimes even harm the scores, although the enhanced images have better perceptual quality.